Research On Anchor-free Algorithms Of Occluded Pedestrian Detection

With the booming development of computer vision,there is a strong demand for the detection and perception of pedestrians in visual sources among various industries.Autonomous driving requires accurate detection of pedestrians to complete the tracking and prediction of their trajectories,and the video surveillance also requires accurate detections in order to provide robust features for person tracking and person re-identification.Occluded pedestrian detection is a difficult subject in pedestrian perception,including the detection of pedestrians occluded by objects or by other pedestrians(dense crowds).Most existing pedestrian detection methods are mainly anchored-based methods,which not only have complicated hyperparameters,complex network structures and slow detection speed,but also result in poor performance in cross-domain scenarios.In recent years,various anchor-free detection algorithms have been proposed to overcome the shortcomings of anchor-based algorithms and achieve detection accuracy comparable to that of anchor-based algorithms,with significant performance improvement especially in cross-domain scenarios.However,there is still a gap between anchor-free methods and the leading anchor-based algorithms in a single domain.In this thesis,we design novel network structures and post-processing method to improve the accuracy of the anchor-free pedestrian detection algorithm for object-obscured pedestrians and crowded pedestrians.The main contents of this thesis include:1.For the problem of unstable and difficult detection of occluded pedestrian features,this thesis proposes an anchor-free pedestrian detection network based on auxiliary supervision.In this thesis,we analyze the reasons for inaccurate detection of occluded pedestrians and design an anchorfree detection network using pedestrian head information as an auxiliary task by using existing and extra manual supervision information.The detection performance of this model is significantly improved over the baseline model in pedestrian detection and crowded pedestrian detection.2.To address the problem that it is difficult to design a cascade structure for abstract features in anchor-free pedestrian detector and to perform iterative optimization of features,this thesis designs an alternating bi-center detection network.This network enables deeper feature interaction and task alternation between additional supervised information to further improve the performance of pedestrian detection.In order to alleviate the detection problem of crowd scenes in pedestrian detection,this thesis further designs "individual" loss and "individual" non-maximum suppression for the above network structure.Based on the "individual" loss guided "individual" non-maximum suppression strategy,the detection of crowd scenes is effectively preserved.This detection algorithm achieves SoTA results in the pedestrian detection dataset,achieves leading detection performance in the crowd human detection dataset,and also achieve best performance in the cross-domain settings,with the detection speed ahead of existing methods.Based on the anchor-free pedestrian detection algorithm,this thesis designs and verifies the gain of the auxiliary task type,especially the alternating learning approach,on the anchor-free pedestrian detection algorithm,and designs a more suitable post-processing strategy,which significantly improves the performance of the baseline anchor-free detector.The model in this thesis enjoys faster speed and leading detection performance,showing the value of theoretical and practical applications.